Rolling bearing fault diagnosis based on the combination of improved deep convolution network and gated recurrent unit

Aiming at the problems of traditional fault diagnosis relying on expert experience, poor generalization ability and complex model training parameters, a new intelligent fault diagnosis model that combines an Improved Deep Convolution Neural Network (IDCNN) and Gated Recurrent Unit (GRU) is proposed. In the process of model building, the original data is expanded by overlapping sampling method to balance the number of different fault signals. Use IDCNN to directly extract features of bearing vibration signals and introduce GRU to mine the timing information hidden in the original signals. At the same time, Global Average Pooling (GAP) is used to replace the fully connected layer to reduce model parameters. Finally, the fault classification is performed through the Softmax layer. The fault recognition accuracy under different working conditions and different SNRs are compared with other three machine learning methods including MLP, FCN and LSTM. The experimental results shows that the proposed method outperforms other three existing methods. Confusion matrix and T-SNE features are also visualized to show the effectiveness ofthe proposed method.